Doiyan/vits-models
0
1import math2import torch3from torch import nn4from torch.nn import functional as F5 6import commons7import modules8import attentions9import monotonic_align10 11from torch.nn import Conv1d, ConvTranspose1d, Conv2d12from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm13from commons import init_weights, get_padding14 15 16class StochasticDurationPredictor(nn.Module):17 def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):18 super().__init__()19 filter_channels = in_channels # it needs to be removed from future version.20 self.in_channels = in_channels21 self.filter_channels = filter_channels22 self.kernel_size = kernel_size23 self.p_dropout = p_dropout24 self.n_flows = n_flows25 self.gin_channels = gin_channels26 27 self.log_flow = modules.Log()28 self.flows = nn.ModuleList()29 self.flows.append(modules.ElementwiseAffine(2))30 for i in range(n_flows):31 self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))32 self.flows.append(modules.Flip())33 34 self.post_pre = nn.Conv1d(1, filter_channels, 1)35 self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)36 self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)37 self.post_flows = nn.ModuleList()38 self.post_flows.append(modules.ElementwiseAffine(2))39 for i in range(4):40 self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))41 self.post_flows.append(modules.Flip())42 43 self.pre = nn.Conv1d(in_channels, filter_channels, 1)44 self.proj = nn.Conv1d(filter_channels, filter_channels, 1)45 self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)46 if gin_channels != 0:47 self.cond = nn.Conv1d(gin_channels, filter_channels, 1)48 49 def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):50 x = torch.detach(x)51 x = self.pre(x)52 if g is not None:53 g = torch.detach(g)54 x = x + self.cond(g)55 x = self.convs(x, x_mask)56 x = self.proj(x) * x_mask57 58 if not reverse:59 flows = self.flows60 assert w is not None61 62 logdet_tot_q = 0 63 h_w = self.post_pre(w)64 h_w = self.post_convs(h_w, x_mask)65 h_w = self.post_proj(h_w) * x_mask66 e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask67 z_q = e_q68 for flow in self.post_flows:69 z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))70 logdet_tot_q += logdet_q71 z_u, z1 = torch.split(z_q, [1, 1], 1) 72 u = torch.sigmoid(z_u) * x_mask73 z0 = (w - u) * x_mask74 logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1,2])75 logq = torch.sum(-0.5 * (math.log(2*math.pi) + (e_q**2)) * x_mask, [1,2]) - logdet_tot_q76 77 logdet_tot = 078 z0, logdet = self.log_flow(z0, x_mask)79 logdet_tot += logdet80 z = torch.cat([z0, z1], 1)81 for flow in flows:82 z, logdet = flow(z, x_mask, g=x, reverse=reverse)83 logdet_tot = logdet_tot + logdet84 nll = torch.sum(0.5 * (math.log(2*math.pi) + (z**2)) * x_mask, [1,2]) - logdet_tot85 return nll + logq # [b]86 else:87 flows = list(reversed(self.flows))88 flows = flows[:-2] + [flows[-1]] # remove a useless vflow89 z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale90 for flow in flows:91 z = flow(z, x_mask, g=x, reverse=reverse)92 z0, z1 = torch.split(z, [1, 1], 1)93 logw = z094 return logw95 96 97class DurationPredictor(nn.Module):98 def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):99 super().__init__()100 101 self.in_channels = in_channels102 self.filter_channels = filter_channels103 self.kernel_size = kernel_size104 self.p_dropout = p_dropout105 self.gin_channels = gin_channels106 107 self.drop = nn.Dropout(p_dropout)108 self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size//2)109 self.norm_1 = modules.LayerNorm(filter_channels)110 self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)111 self.norm_2 = modules.LayerNorm(filter_channels)112 self.proj = nn.Conv1d(filter_channels, 1, 1)113 114 if gin_channels != 0:115 self.cond = nn.Conv1d(gin_channels, in_channels, 1)116 117 def forward(self, x, x_mask, g=None):118 x = torch.detach(x)119 if g is not None:120 g = torch.detach(g)121 x = x + self.cond(g)122 x = self.conv_1(x * x_mask)123 x = torch.relu(x)124 x = self.norm_1(x)125 x = self.drop(x)126 x = self.conv_2(x * x_mask)127 x = torch.relu(x)128 x = self.norm_2(x)129 x = self.drop(x)130 x = self.proj(x * x_mask)131 return x * x_mask132 133 134class TextEncoder(nn.Module):135 def __init__(self,136 n_vocab,137 out_channels,138 hidden_channels,139 filter_channels,140 n_heads,141 n_layers,142 kernel_size,143 p_dropout):144 super().__init__()145 self.n_vocab = n_vocab146 self.out_channels = out_channels147 self.hidden_channels = hidden_channels148 self.filter_channels = filter_channels149 self.n_heads = n_heads150 self.n_layers = n_layers151 self.kernel_size = kernel_size152 self.p_dropout = p_dropout153 154 self.emb = nn.Embedding(n_vocab, hidden_channels)155 nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)156 157 self.encoder = attentions.Encoder(158 hidden_channels,159 filter_channels,160 n_heads,161 n_layers,162 kernel_size,163 p_dropout)164 self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)165 166 def forward(self, x, x_lengths):167 x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]168 x = torch.transpose(x, 1, -1) # [b, h, t]169 x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)170 171 x = self.encoder(x * x_mask, x_mask)172 stats = self.proj(x) * x_mask173 174 m, logs = torch.split(stats, self.out_channels, dim=1)175 return x, m, logs, x_mask176 177 178class ResidualCouplingBlock(nn.Module):179 def __init__(self,180 channels,181 hidden_channels,182 kernel_size,183 dilation_rate,184 n_layers,185 n_flows=4,186 gin_channels=0):187 super().__init__()188 self.channels = channels189 self.hidden_channels = hidden_channels190 self.kernel_size = kernel_size191 self.dilation_rate = dilation_rate192 self.n_layers = n_layers193 self.n_flows = n_flows194 self.gin_channels = gin_channels195 196 self.flows = nn.ModuleList()197 for i in range(n_flows):198 self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))199 self.flows.append(modules.Flip())200 201 def forward(self, x, x_mask, g=None, reverse=False):202 if not reverse:203 for flow in self.flows:204 x, _ = flow(x, x_mask, g=g, reverse=reverse)205 else:206 for flow in reversed(self.flows):207 x = flow(x, x_mask, g=g, reverse=reverse)208 return x209 210 211class PosteriorEncoder(nn.Module):212 def __init__(self,213 in_channels,214 out_channels,215 hidden_channels,216 kernel_size,217 dilation_rate,218 n_layers,219 gin_channels=0):220 super().__init__()221 self.in_channels = in_channels222 self.out_channels = out_channels223 self.hidden_channels = hidden_channels224 self.kernel_size = kernel_size225 self.dilation_rate = dilation_rate226 self.n_layers = n_layers227 self.gin_channels = gin_channels228 229 self.pre = nn.Conv1d(in_channels, hidden_channels, 1)230 self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)231 self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)232 233 def forward(self, x, x_lengths, g=None):234 x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)235 x = self.pre(x) * x_mask236 x = self.enc(x, x_mask, g=g)237 stats = self.proj(x) * x_mask238 m, logs = torch.split(stats, self.out_channels, dim=1)239 z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask240 return z, m, logs, x_mask241 242 243class Generator(torch.nn.Module):244 def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):245 super(Generator, self).__init__()246 self.num_kernels = len(resblock_kernel_sizes)247 self.num_upsamples = len(upsample_rates)248 self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)249 resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2250 251 self.ups = nn.ModuleList()252 for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):253 self.ups.append(weight_norm(254 ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),255 k, u, padding=(k-u)//2)))256 257 self.resblocks = nn.ModuleList()258 for i in range(len(self.ups)):259 ch = upsample_initial_channel//(2**(i+1))260 for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):261 self.resblocks.append(resblock(ch, k, d))262 263 self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)264 self.ups.apply(init_weights)265 266 if gin_channels != 0:267 self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)268 269 def forward(self, x, g=None):270 x = self.conv_pre(x)271 if g is not None:272 x = x + self.cond(g)273 274 for i in range(self.num_upsamples):275 x = F.leaky_relu(x, modules.LRELU_SLOPE)276 x = self.ups[i](x)277 xs = None278 for j in range(self.num_kernels):279 if xs is None:280 xs = self.resblocks[i*self.num_kernels+j](x)281 else:282 xs += self.resblocks[i*self.num_kernels+j](x)283 x = xs / self.num_kernels284 x = F.leaky_relu(x)285 x = self.conv_post(x)286 x = torch.tanh(x)287 288 return x289 290 def remove_weight_norm(self):291 print('Removing weight norm...')292 for l in self.ups:293 remove_weight_norm(l)294 for l in self.resblocks:295 l.remove_weight_norm()296 297 298class DiscriminatorP(torch.nn.Module):299 def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):300 super(DiscriminatorP, self).__init__()301 self.period = period302 self.use_spectral_norm = use_spectral_norm303 norm_f = weight_norm if use_spectral_norm == False else spectral_norm304 self.convs = nn.ModuleList([305 norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),306 norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),307 norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),308 norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),309 norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),310 ])311 self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))312 313 def forward(self, x):314 fmap = []315 316 # 1d to 2d317 b, c, t = x.shape318 if t % self.period != 0: # pad first319 n_pad = self.period - (t % self.period)320 x = F.pad(x, (0, n_pad), "reflect")321 t = t + n_pad322 x = x.view(b, c, t // self.period, self.period)323 324 for l in self.convs:325 x = l(x)326 x = F.leaky_relu(x, modules.LRELU_SLOPE)327 fmap.append(x)328 x = self.conv_post(x)329 fmap.append(x)330 x = torch.flatten(x, 1, -1)331 332 return x, fmap333 334 335class DiscriminatorS(torch.nn.Module):336 def __init__(self, use_spectral_norm=False):337 super(DiscriminatorS, self).__init__()338 norm_f = weight_norm if use_spectral_norm == False else spectral_norm339 self.convs = nn.ModuleList([340 norm_f(Conv1d(1, 16, 15, 1, padding=7)),341 norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),342 norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),343 norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),344 norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),345 norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),346 ])347 self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))348 349 def forward(self, x):350 fmap = []351 352 for l in self.convs:353 x = l(x)354 x = F.leaky_relu(x, modules.LRELU_SLOPE)355 fmap.append(x)356 x = self.conv_post(x)357 fmap.append(x)358 x = torch.flatten(x, 1, -1)359 360 return x, fmap361 362 363class MultiPeriodDiscriminator(torch.nn.Module):364 def __init__(self, use_spectral_norm=False):365 super(MultiPeriodDiscriminator, self).__init__()366 periods = [2,3,5,7,11]367 368 discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]369 discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]370 self.discriminators = nn.ModuleList(discs)371 372 def forward(self, y, y_hat):373 y_d_rs = []374 y_d_gs = []375 fmap_rs = []376 fmap_gs = []377 for i, d in enumerate(self.discriminators):378 y_d_r, fmap_r = d(y)379 y_d_g, fmap_g = d(y_hat)380 y_d_rs.append(y_d_r)381 y_d_gs.append(y_d_g)382 fmap_rs.append(fmap_r)383 fmap_gs.append(fmap_g)384 385 return y_d_rs, y_d_gs, fmap_rs, fmap_gs386 387 388 389class SynthesizerTrn(nn.Module):390 """391 Synthesizer for Training392 """393 394 def __init__(self, 395 n_vocab,396 spec_channels,397 segment_size,398 inter_channels,399 hidden_channels,400 filter_channels,401 n_heads,402 n_layers,403 kernel_size,404 p_dropout,405 resblock, 406 resblock_kernel_sizes, 407 resblock_dilation_sizes, 408 upsample_rates, 409 upsample_initial_channel, 410 upsample_kernel_sizes,411 n_speakers=0,412 gin_channels=0,413 use_sdp=True,414 **kwargs):415 416 super().__init__()417 self.n_vocab = n_vocab418 self.spec_channels = spec_channels419 self.inter_channels = inter_channels420 self.hidden_channels = hidden_channels421 self.filter_channels = filter_channels422 self.n_heads = n_heads423 self.n_layers = n_layers424 self.kernel_size = kernel_size425 self.p_dropout = p_dropout426 self.resblock = resblock427 self.resblock_kernel_sizes = resblock_kernel_sizes428 self.resblock_dilation_sizes = resblock_dilation_sizes429 self.upsample_rates = upsample_rates430 self.upsample_initial_channel = upsample_initial_channel431 self.upsample_kernel_sizes = upsample_kernel_sizes432 self.segment_size = segment_size433 self.n_speakers = n_speakers434 self.gin_channels = gin_channels435 436 self.use_sdp = use_sdp437 438 self.enc_p = TextEncoder(n_vocab,439 inter_channels,440 hidden_channels,441 filter_channels,442 n_heads,443 n_layers,444 kernel_size,445 p_dropout)446 self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)447 self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)448 self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)449 450 if use_sdp:451 self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)452 else:453 self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)454 455 if n_speakers > 1:456 self.emb_g = nn.Embedding(n_speakers, gin_channels)457 458 def forward(self, x, x_lengths, y, y_lengths, sid=None):459 460 x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)461 if self.n_speakers > 0:462 g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]463 else:464 g = None465 466 z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)467 z_p = self.flow(z, y_mask, g=g)468 469 with torch.no_grad():470 # negative cross-entropy471 s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]472 neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]473 neg_cent2 = torch.matmul(-0.5 * (z_p ** 2).transpose(1, 2), s_p_sq_r) # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]474 neg_cent3 = torch.matmul(z_p.transpose(1, 2), (m_p * s_p_sq_r)) # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]475 neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]476 neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4477 478 attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)479 attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()480 481 w = attn.sum(2)482 if self.use_sdp:483 l_length = self.dp(x, x_mask, w, g=g)484 l_length = l_length / torch.sum(x_mask)485 else:486 logw_ = torch.log(w + 1e-6) * x_mask487 logw = self.dp(x, x_mask, g=g)488 l_length = torch.sum((logw - logw_)**2, [1,2]) / torch.sum(x_mask) # for averaging 489 490 # expand prior491 m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)492 logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)493 494 z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)495 o = self.dec(z_slice, g=g)496 return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)497 498 def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):499 x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)500 if self.n_speakers > 0:501 g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]502 else:503 g = None504 505 if self.use_sdp:506 logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)507 else:508 logw = self.dp(x, x_mask, g=g)509 w = torch.exp(logw) * x_mask * length_scale510 w_ceil = torch.ceil(w)511 y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()512 y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)513 attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)514 attn = commons.generate_path(w_ceil, attn_mask)515 516 m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']517 logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']518 519 z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale520 z = self.flow(z_p, y_mask, g=g, reverse=True)521 o = self.dec((z * y_mask)[:,:,:max_len], g=g)522 return o, attn, y_mask, (z, z_p, m_p, logs_p)523 524 def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):525 assert self.n_speakers > 0, "n_speakers have to be larger than 0."526 g_src = self.emb_g(sid_src).unsqueeze(-1)527 g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)528 z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)529 z_p = self.flow(z, y_mask, g=g_src)530 z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)531 o_hat = self.dec(z_hat * y_mask, g=g_tgt)532 return o_hat, y_mask, (z, z_p, z_hat)533 534